{"id":"a4190dfc-32ed-4f16-916e-e17fb377cd7e","arxiv_id":"2412.10878","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An adaptive mixed-resolution gradient quantizer plus uplink power control reduces federated learning communication overhead by over 90% while keeping test accuracy close to full-precision FL on CIFAR and Fashion-MNIST.","lead":"Federated learning trains a shared model across phones or sensors without moving private data, but sending model updates over wireless networks is slow. This paper mixes two ideas, using few bits for unimportant gradient values and adjusting transmit power to keep the slowest user from dragging down each round, and reports up to 93% less communication traffic on standard image benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The mixed-resolution bitstream of Eq. (7)-(8) is not uniquely decodable without a per-coordinate mask; including the mask roughly halves the claimed Table III overhead reduction, so the benchmark accuracy gain must be recomputed.","rationale":"I read the paper as a communication-system plus FL contribution whose central quantitative claim is the 93% overhead reduction and the 10% accuracy advantage under a fixed latency budget. Both rest on the bit-count formula in Section II-C. That formula silently assumes the receiver can distinguish one-bit sign codewords from b_j-bit quantized codewords; in any serialized transmission this requires a mask or equivalent, which the paper does not provide. This is not a question of whether gradients are sparse; it is a property of the encoding. The reader's sparsity concern is related, since if s is large the savings shrink, but it presumes the coding is realizable. I therefore keep the verdict conditional: the scheme is promising, but the paper must specify the side information, rerun Table III with corrected bits, and re-examine the 10% accuracy claim. I do not see a reason to reject outright, as a mask is a straightforward protocol addition and the full-precision comparison may partially survive; I also do not see fraud, only a missing protocol detail with quantitative consequences.","tokens_in":10775,"tokens_out":15619,"duration_ms":155792,"concrete_test":"Write a decoder for the exact bitstream produced by Eq. (7)-(8) with b_j=4, lambda=0.2, d=3, delta=[1,0.1,-0.05]; show that more than one parse exists, e.g., treating the first symbol as high-resolution versus low-resolution yields different reconstructed vectors. Then recompute the Table III latency/accuracy comparison using the corrected overhead d(2+(b_j-1)s)+32, i.e., one mask bit per element. If the accuracy gap versus AQUILA/LAQ falls below 10%, the abstract's benchmark claim is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is not data sparsity but the encoding itself. In Section II-C, low-resolution entries are coded as a single sign bit (0 or 1) and high-resolution entries as a b_j-bit uniform-quantizer output, and the total bit count is d((1-s)+b_j s)+32 with no side information. These variable-length codewords are not uniquely decodable: the one-bit codewords are prefixes of every b_j-bit codeword, so a receiver handed the serialized bitstream cannot determine which coordinates use 1 bit and which use b_j bits. No mask, index list, or prefix code is specified anywhere in Eq. (7)-(8), Algorithm 1, or Section II-C. The protocol as written is therefore not implementable. If a d-bit mask is added, the true per-vector cost becomes d(2+(b_j-1)s)+32 bits. For the Table III configuration (b_j=4, lambda=0.4, s=0.044%), this doubles the payload from about 1.001 to about 2.001 bits/element and changes the claimed 75% overhead reduction versus AQUILA/LAQ into roughly 50%. Since Tmax under the fixed 3 s latency budget is determined by these bits, the reported Tmax and the 10% accuracy margin in the abstract would have to be re-derived. The sparsity assumption the reader flagged may also fail on other architectures, but this coding ambiguity is more immediate: it affects every reported overhead number, not just the robustness of Table II.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive mixed-resolution quantization scheme for federated learning over cell-free massive MIMO: gradient entries below a magnitude-ratio threshold are encoded with a single sign bit, while the remaining entries are uniformly quantized with b_j bits, and uplink powers are optimized by bisection over an epigraph-form min-max problem to mitigate stragglers. The authors claim a convergence bound for FL with local AdaGrad updates (Proposition 1), overhead reductions of about 93–96% relative to full-precision FL with comparable accuracy on CIFAR-10, CIFAR-100, and Fashion-MNIST, and a 75% overhead reduction and 10% accuracy gain over AQUILA, LAQ, and Top-q under a fixed latency budget (Tables II–III).","tokens_in":11150,"tokens_out":12870,"duration_ms":121873,"significance":"The problem is timely and the system model is detailed; the power-control reformulation in Eq. (14) is elegant, and the bisection/LP solution is a sensible approach. The paper also makes a concrete attempt to prove a quantization-error bound in Lemma 1. If the central claims held, the work would be a useful contribution to communication-efficient FL over cell-free networks. However, the protocol as written is not implementable because the mixed-resolution bitstream is not uniquely decodable, and the main theoretical bound has a serious proof gap. Both issues are load-bearing for the headline overhead and accuracy claims, so the manuscript needs substantial revision.","major_comments":[{"comment":"The serialized bitstream is not uniquely decodable. Low-resolution entries are coded as a single bit (0 or 1), high-resolution entries as a b_j-bit uniform-quantizer codeword, and the total is counted as d((1-s)+b_j s)+32 bits with no mask, index list, or prefix code. There is no specified restriction preventing a one-bit codeword from being a prefix of a b_j-bit codeword, so a receiver cannot determine which coordinates use which length. If a d-bit mask is added, the true per-vector cost becomes d(2+(b_j-1)s)+32 bits; for the Table III configuration (b_j=4, s=0.044%) this changes the claimed overhead reduction versus LAQ/AQUILA from about 75% to about 50%, and T_max and the reported accuracy margin must be recomputed. This issue affects every overhead and latency number in the paper.","section":"Section II-C, Eq. (7)–(8)"},{"comment":"The proof of Lemma 1 bounds the signed difference [ε_j^t]_i = [δw_j^t]_i - cδw_j^q,t/2, not its absolute value. For a low-resolution entry of magnitude x ∈ [0, λ∥δw_j^t∥∞), the actual reconstruction error is |x - cδw_j^q,t/2|, whose supremum is cδw_j^q,t/2, which can approach ∥δw_j^t∥∞/2 when the smallest high-resolution element q_t is close to ∥δw_j^t∥∞. For example, with λ=0.1, ∥δw_j^t∥∞=1, one high-resolution entry of magnitude 1 and one low-resolution entry of magnitude 0.05, q_t≈1 and the reconstruction error for the low entry is about 0.45, far above the claimed c_j≈0.05. Therefore Lemma 1 does not follow, and the Kc_max^2 term in Proposition 1, Eq. (10), is unsupported.","section":"Appendix A, Lemma 1 and Eq. (17)"},{"comment":"The proof is deferred to Reddi et al. [22] with 'details omitted due to limited space.' This is not a direct substitution: the quantized update is not unbiased in general, and the stated conditions I–II of [22] are not verified for this quantizer. Moreover, as printed, Φ contains a term proportional to T (2LT σ̄^2), so the right-hand side of Eq. (10) tends to a nonzero constant as T grows; the claimed convergence to a stationary point does not follow from the displayed expression. A self-contained proof or a precise reduction that checks all assumptions is required.","section":"Proposition 1, Eq. (10)"},{"comment":"The central empirical claims—matching full-precision FL accuracy and the 10% accuracy margin over benchmarks—are based on single-point accuracy numbers with no reported seeds, variance, or error bars. Table III's T_max values (e.g., 27 vs 17 for our power control vs LAQ) are the basis for the latency-budget comparison, but without repeated runs it is not possible to assess whether the differences are statistically meaningful. The absence of code further hampers reproducibility. These tables should be regenerated with multiple seeds and confidence intervals, and the claims adjusted if the side-information overhead from Major Comment 1 changes the bit counts.","section":"Section IV, Tables II–III"}],"minor_comments":[{"comment":"The local update is printed as w_j^{h,t} ← −w_j^{h−1,t} − α ...; this appears to be a sign typo. The standard AdaGrad local update should be w_j^{h,t} ← w_j^{h−1,t} − α ...; as printed, the local iterate alternates in sign and is inconsistent with the rest of the paper.","section":"Eq. (2)"},{"comment":"The notation b_j^t is used both for the d-dimensional bit vector in Eq. (7) and for the scalar total number of bits in Section II-C and Algorithm 1; please use different symbols (e.g., β_j^t for the vector) to avoid ambiguity.","section":"Section II-C and Algorithm 1"},{"comment":"The abstract claims 'at least 93%' overhead reduction, while Section IV states 'at least r̄=96%' for Table II; these numbers should be reconciled and the parameter configurations (λ, b_j) behind each claim should be stated explicitly.","section":"Abstract and Section IV"},{"comment":"Line 11 says users send b_j^t with b_j^p,t = ⌈log2(b_j^t)⌉ bits, but b_j^t is already defined as the scalar bit count; the purpose of this transmission and the distinction between the bit count and its encoding should be clarified.","section":"Algorithm 1, lines 10–11"},{"comment":"The method's overhead savings rest on the sparsity of local gradients invoked in Eq. (6)–(7), but only one CNN architecture is tested. The paper should at least discuss sensitivity to architecture and dataset, or add a robustness experiment, since the measured s values in Table II may not transfer to other models.","section":"Section IV and Table II"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript has two load-bearing technical problems—the non-decodable mixed-resolution bitstream and the incorrect Lemma 1 bound—that preclude acceptance in the current form. The core power-control formulation and the general idea remain plausible, and both issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The empirical tables also need multiple seeds before the headline accuracy and latency claims can be evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know up front. The adaptive mixed-resolution idea is genuine and natural: keep one sign bit for small gradient entries, use b_j bits for the few large ones, and co-optimize transmit powers with the resulting bit counts to tame stragglers in cell-free massive MIMO. The experimental story is plausible, and accuracy is close to full-precision FL on three datasets. But the bit-counting is wrong as written. Eq. (7)-(8) and the total bit formula omit the mask needed to tell the receiver which entries are 1-bit signs and which are b_j-bit codewords. Without a mask or prefix code, the serialized vector is not uniquely decodable. Adding a d-bit mask changes the per-vector cost to d(2+(b_j-1)s)+32, which for their Table III setting (s=0.044%, b_j=4) doubles the payload and cuts the claimed 75% overhead reduction over LAQ/AQUILA to about 50%. The 93% figure in the abstract is likewise affected. That does not sink the core idea, but it means the headline numbers need to be recomputed and the accuracy comparison under a fixed latency budget re-run.\n\nWhat the paper does well: Lemma 1 is a direct, checkable error bound; the power-control problem is a standard epigraph+bisection formulation and is solved exactly; the three-dataset comparison with AQUILA, LAQ, and Top-q is the right benchmark set. The convergence claim in Proposition 1 is borrowed from Reddi et al. with details omitted; that is acceptable for a workshop-grade argument but it is not a proof, and I would want either the full argument or an explicit statement that convergence is empirical. Tables II-III have no error bars, no seeds, and no code, so the numerical claims are single-run evidence. That is a normal weakness for this subfield, but it matters here because the overhead fractions s are what drive the savings.\n\nThe reader's note flagged sparsity as the weakest assumption. I disagree: the encoding ambiguity is more immediate, because it touches every overhead number. The sparsity assumption is at least empirically supported by Table II.\n\nBottom line: this deserves a serious referee. The missing mask is a fixable error, not a broken concept. My recommendation: engage, but require the authors to add the mask (or use a prefix code), recompute all overhead and latency numbers, and release code or at least seeds and error bars. Then it could be a solid systems paper.","headline":"The mixed-resolution quantizer is a sensible idea, but the bitstream omits a mask, so the headline overhead numbers are off by about a factor of two; still worth a careful revision.","tokens_in":11647,"tokens_out":3620,"would_cite":false,"duration_ms":32794,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Federated learning over cell-free networks can cut uplink communication overhead by at least 93% using adaptive mixed-resolution quantization that spends full bits only on the largest gradient entries, while keeping test accuracy close to…","keywords":["federated learning","cell-free massive MIMO","adaptive quantization","uplink power control","straggler mitigation","communication efficiency","gradient sparsity","latency"],"falsifier":"Run the same algorithm on a model class whose gradients are known to be dense, such as a small ResNet or a transformer on a language task, and measure the average high-resolution fraction $s$ and the accuracy gap to full-precision FL; if $s$ exceeds roughly 5-10%, the claimed at-least-93% overhead reduction cannot hold, and the convergence bound shows the gap grows with $K c_{max}^2$.","tokens_in":1778,"feed_emoji":"📉","tokens_out":2050,"duration_ms":65986,"temperature":0.7,"pith_summary":"The paper tries to show that the uplink bottleneck of federated learning over cell-free massive MIMO, the latency of sending full-precision gradient vectors from many users, can be broken by recognizing that most gradient entries are nearly zero. It proposes an element-wise quantizer that gives a few large entries high-resolution uniform quantization and compresses all remaining entries to a single sign bit, so the number of bits each user sends varies per iteration. It then sets uplink powers to maximize the smallest rate-per-bit ratio, which prevents the slowest user from dictating the round time. Across CIFAR-10, CIFAR-100, and Fashion-MNIST, the scheme reportedly keeps test accuracy close to classic FL while reducing communication overhead by at least 93%, and beats AQUILA, Top-q, and LAQ by 10% accuracy within a fixed latency budget.","feed_headline":"Mixed-resolution gradients cut FL uplink traffic by 93%","feed_subtitle":"Adaptive quantization keeps accuracy close to full precision while power control stops stragglers.","key_machinery":"The load-bearing object is the adaptive mixed-resolution quantization rule in (6)-(7) together with the rate-per-bit objective in (13). The quantizer exploits the empirically observed sparsity of local gradients: the fraction $s_j^t$ of entries above the magnitude ratio threshold $\\lambda_j$ is typically below 2% in the experiments, so roughly 98% of entries are transmitted as a single sign bit. The power control solves the minimax latency problem by bisection over the lowest rate-per-bit ratio and a linear program for the user powers, explicitly balancing the number of bits each user owes against the rate the cell-free network can give that user.","core_discovery":"The central claim is that a threshold-based mixed-resolution quantizer defined by $\\lambda_j$ relative to the infinity norm of each local gradient update preserves federated learning convergence while making uplink traffic almost negligible. For entries with $|[\\delta w_j^t]_i| / \\|\\delta w_j^t\\|_\\infty < \\lambda_j$, only the sign is sent; the few entries at or above the threshold are uniformly quantized with $b_j$ bits, and the reconstruction sets the low-resolution magnitudes to half the quantized smallest high-resolution value. Coupled with a min-max power control that maximizes the smallest rate-per-bit ratio via bisection and linear programming, the method lets more FL iterations fit in a given latency budget, which is why it gains 10% accuracy over fixed-quantization benchmarks under the same total time.","pith_inferences":["The sparsity assumption is the hinge: the reported 93-96% savings reflect $s \\approx 0.7\\%\\text{--}1.8\\%$ on these CNN/dataset combinations. For architectures with denser gradients (residual networks, transformers, or batch-normalized models), the savings would shrink and the accuracy gap could widen; that is a testable extension, not a claim in the paper.","The low-resolution entries are reconstructed at half the smallest high-resolution quantized value, which injects a systematic magnitude bias toward zero. A bias-correction or dithering variant could push accuracy closer to full precision while keeping the same bit budget.","The per-user threshold $\\lambda_j$ is fixed in the experiments; adapting it to channel quality (weaker users could send fewer high-resolution entries) would trade convergence per round against latency exactly along the rate-per-bit curve the paper already defines.","The same mixed-resolution idea could be applied to the downlink broadcast of the global model, where the paper assumes error-free transmission; the bit-saving logic does not depend on the uplink-only setting."],"forward_implications":["Uplink communication per FL round no longer scales with 32 bits times model dimension; with $s \\approx 1\\%$, the per-user payload is dominated by one bit per parameter, reducing overhead by at least 93% on the tested datasets.","Because the bit count $b_j^t$ adapts per user and iteration, the power control can equalize finish times, so iteration latency is set by the maximum rate-per-bit rather than the worst channel user.","Under a fixed total latency budget, the method completes more global rounds (up to $T_{max}=27$ versus 16 for AQUILA in the reported setting), which is the direct mechanism for the 10% accuracy gain over benchmarks.","The convergence bound (10) shows the quantization error enters only through $K c_{max}^2(2L + T \\bar\\sigma^2)$, so as long as the per-entry error bound $c_j$ stays small, AdaGrad FL converges at the same order as unquantized adaptive federated optimization."],"supporting_citations":[{"why":"Supplies the AdaGrad local-update rule and the convergence assumptions that Proposition 1 extends to the quantized setting.","marker":"[22]"},{"why":"Supplies the cell-free massive MIMO system model, pilot assignment, and the max-sum rate benchmark used in comparisons.","marker":"[2]"},{"why":"Provides the achievable uplink rate expression that the paper uses to define each user's latency.","marker":"[25]"},{"why":"Provides the uplink process detail for FL over cell-free massive MIMO that Algorithm 1 builds on.","marker":"[24]"},{"why":"AQUILA is an adaptive quantization baseline compared against the proposed method under the same power control schemes.","marker":"[8]"},{"why":"LAQ is a lazily aggregated quantized gradient baseline compared in the latency-budget experiments.","marker":"[9]"},{"why":"Supplies the Top-q gradient sparsification baseline, where only the q largest entries are transmitted.","marker":"[20]"},{"why":"Provides the Dinkelbach power control benchmark used to evaluate the proposed power control scheme.","marker":"[21]"},{"why":"Supports the premise that local gradients are sparse, which motivates the mixed-resolution quantization design.","marker":"[17]"}],"fun_headline_variants":["Cell-free FL: adaptive quantization cuts uplink 93%","Mixed-resolution gradients cut FL uplink by 93%","Adaptive quantization and power control slash FL overhead","Federated learning gets 93% leaner with cell-free adaptive quantization","93% less uplink data in federated learning via adaptive bits"],"cache_read_input_tokens":13696,"weakest_assumption_plain":"The entire overhead reduction rests on the empirical regularity that most local gradient entries are so close to zero that replacing their magnitude with one sign bit does not meaningfully hurt convergence; the paper measures this sparsity ($s \\approx 0.7\\%\\text{--}1.8\\%$) on three datasets but gives no theoretical guarantee or cross-architecture evidence that it always holds.","fun_headline_variants_meta":{"raw":{"variants":["Cell-free FL: adaptive quantization cuts uplink 93%","Mixed-resolution gradients cut FL uplink by 93%","Adaptive quantization and power control slash FL overhead","Federated learning gets 93% leaner with cell-free adaptive quantization","93% less uplink data in federated learning via adaptive bits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2512,"prompt_tokens":978,"completion_tokens":1534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":1448}},"tokens_in":594,"tokens_out":1534,"duration_ms":12954,"temperature":1.0,"reasoning_tokens":1448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:31:34.316675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same algorithm on a model class whose gradients are known to be dense, such as a small ResNet or a transformer on a language task, and measure the average high-resolution fraction $s$ and the accuracy gap to full-precision FL; if $s$ exceeds roughly 5-10%, the claimed at-least-93% overhead reduction cannot hold, and the convergence bound shows the gap grows with $K c_{max}^2$.","supporting_citations":[{"cited_title":"Foundations of user-centric cell-free Massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the cell-free massive MIMO system model, pilot assignment, and the max-sum rate benchmark used in comparisons."},{"cited_title":"Cell-free Massive MIMO versus small cells,","cited_arxiv_id":null,"evidence_quote":"Provides the achievable uplink rate expression that the paper uses to define each user's latency."},{"cited_title":"Joint energy and latency optimization in Federated Learning over cell-free massive MIMO networks,","cited_arxiv_id":null,"evidence_quote":"Provides the uplink process detail for FL over cell-free massive MIMO that Algorithm 1 builds on."},{"cited_title":"AQUILA: Communication efficient Federated Learning with adaptive quantization in device selection strategy,","cited_arxiv_id":null,"evidence_quote":"AQUILA is an adaptive quantization baseline compared against the proposed method under the same power control schemes."},{"cited_title":"Lazily Aggregated Quantized Gradient (LAQ) innovation for communication-efficient Federated Learning,","cited_arxiv_id":null,"evidence_quote":"LAQ is a lazily aggregated quantized gradient baseline compared in the latency-budget experiments."},{"cited_title":"Gradient sparsification for communication-efficient distributed optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the Top-q gradient sparsification baseline, where only the q largest entries are transmitted."},{"cited_title":"Energy-efficient power control: A look at 5G wireless technologies,","cited_arxiv_id":null,"evidence_quote":"Provides the Dinkelbach power control benchmark used to evaluate the proposed power control scheme."},{"cited_title":"SAFARI: Sparsity-enabled Federated Learning with limited and unreliable communications,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that local gradients are sparse, which motivates the mixed-resolution quantization design."}],"review_version":1}